Agent Router unifies model calls but can't route SDK-built agents

The author set out to test Agent Router (formerly Envoy AI Gateway) after reading about it and not fully understanding its role. The tool is an open-source AI gateway that accepts LLM requests from applications and routes them to the appropriate provider or model based on configuration. To try it, the author installed the aigw command-line tool, wrote a YAML file defining routes for Amazon Bedrock's Claude and GPT models, started the gateway with aigw run agent-router.yaml, and then sent test requests to a single endpoint at localhost:1975.

The test showed that Agent Router can collapse provider-specific API differences into one POST /v1/chat/completions endpoint, so the calling application does not need to branch its code by service. But the author's actual motivation was different: they wanted to route AI agents running inside a web application. While the gateway seems well suited to routing API requests from web applications to various providers and models, routing agents built with various SDKs appeared difficult, and the author abandoned that goal for the time being. The write-up thus draws a clear line between what the tool does well today and what it does not yet handle.